CutClean: Neural Network Pruning for Privacy-Preserving Inference

2026-08-13

AI Summary

This study addresses the issue of sensitive attribute privacy leakage during neural network inference by proposing CutClean, a privacy-aware pruning method. The approach innovatively introduces an auxiliary linear privacy head to quantify private information flow and integrates progressive pruning with sparsity training to precisely eliminate privacy-related features during model compression. Experimental results demonstrate that CutClean significantly enhances model sparsity while effectively mitigating privacy leakage risks and maintaining high target classification accuracy. Consequently, this method achieves a favorable trade-off among privacy protection, model lightweighting, and task performance, offering a robust solution for deploying secure and efficient neural networks in privacy-sensitive scenarios.

Abstract

Neural networks are increasingly deployed in high-stakes applications with growing privacy leakage concerns. We show that this privacy leakage can occur even in the absence of representation imbalances that lead to traditional dataset biases. This poses significant privacy risks when deploying models that process sensitive attributes. In this context, we propose CutClean, a privacy-aware pruning method that allows to reduce privacy information flow through the network, while increasing its sparsity. Our approach employs auxiliary linear privacy heads placed at each network's block to quantify information leakage, and further applies increasing levels of sparsity to remove the private attribute leakage, measured in terms of the accuracy of the privacy head attached to the last block. Experiments on synthetic and real-world datasets demonstrate that our approach effectively minimizes private information flow while achieving high sparsity rates and preserving classification target accuracy.

Problem

• Privacy Leakage
• Neural Network Pruning
• Sensitive Attributes
• Privacy-Preserving Inference

Innovation

• Privacy-aware Pruning
• Auxiliary Privacy Heads
• Information Leakage Quantification
• Incremental Sparsity
• Privacy-Preserving Inference